Tian Guo

Yiqin Zhao, Sean Fanello, Tian Guo

Lighting understanding plays an important role in virtual object composition, including mobile augmented reality (AR) applications. Prior work often targets recovering lighting from the physical environment to support photorealistic AR rendering. Because the common workflow is to use a back-facing camera to capture the physical world for overlaying virtual objects, we refer to this usage pattern as back-facing AR. However, existing methods often fall short in supporting emerging front-facing mobile AR applications, e.g., virtual try-on where a user leverages a front-facing camera to explore the effect of various products (e.g., glasses or hats) of different styles. This lack of support can be attributed to the unique challenges of obtaining 360$^\circ$ HDR environment maps, an ideal format of lighting representation, from the front-facing camera and existing techniques. In this paper, we propose to leverage dual-camera streaming to generate a high-quality environment map by combining multi-view lighting reconstruction and parametric directional lighting estimation. Our preliminary results show improved rendering quality using a dual-camera setup for front-facing AR compared to a commercial solution.

Yiqin Zhao, Chongyang Ma, Haibin Huang et al.

An accurate understanding of omnidirectional environment lighting is crucial for high-quality virtual object rendering in mobile augmented reality (AR). In particular, to support reflective rendering, existing methods have leveraged deep learning models to estimate or have used physical light probes to capture physical lighting, typically represented in the form of an environment map. However, these methods often fail to provide visually coherent details or require additional setups. For example, the commercial framework ARKit uses a convolutional neural network that can generate realistic environment maps; however the corresponding reflective rendering might not match the physical environments. In this work, we present the design and implementation of a lighting reconstruction framework called LitAR that enables realistic and visually-coherent rendering. LitAR addresses several challenges of supporting lighting information for mobile AR. First, to address the spatial variance problem, LitAR uses two-field lighting reconstruction to divide the lighting reconstruction task into the spatial variance-aware near-field reconstruction and the directional-aware far-field reconstruction. The corresponding environment map allows reflective rendering with correct color tones. Second, LitAR uses two noise-tolerant data capturing policies to ensure data quality, namely guided bootstrapped movement and motion-based automatic capturing. Third, to handle the mismatch between the mobile computation capability and the high computation requirement of lighting reconstruction, LitAR employs two novel real-time environment map rendering techniques called multi-resolution projection and anchor extrapolation. These two techniques effectively remove the need of time-consuming mesh reconstruction while maintaining visual quality.

Yiyang Zhao, Tian Guo

This work presents a novel approach to neural architecture search (NAS) that aims to reduce energy costs and increase carbon efficiency during the model design process. The proposed framework, called carbon-efficient NAS (CE-NAS), consists of NAS evaluation algorithms with different energy requirements, a multi-objective optimizer, and a heuristic GPU allocation strategy. CE-NAS dynamically balances energy-efficient sampling and energy-consuming evaluation tasks based on current carbon emissions. Using a recent NAS benchmark dataset and two carbon traces, our trace-driven simulations demonstrate that CE-NAS achieves better carbon and search efficiency than the three baselines.

Ashkan Ganj, Yiqin Zhao, Hang Su et al.

Metric depth estimation plays an important role in mobile augmented reality (AR). With accurate metric depth, we can achieve more realistic user interactions such as object placement and occlusion detection. While specialized hardware like LiDAR demonstrates its promise, its restricted availability, i.e., only on selected high-end mobile devices, and performance limitations such as range and sensitivity to the environment, make it less ideal. Monocular depth estimation, on the other hand, relies solely on mobile cameras, which are ubiquitous, making it a promising alternative for mobile AR. In this paper, we investigate the challenges and opportunities of achieving accurate metric depth estimation in mobile AR. We tested four different state-of-the-art monocular depth estimation models on a newly introduced dataset (ARKitScenes) and identified three types of challenges: hard-ware, data, and model related challenges. Furthermore, our research provides promising future directions to explore and solve those challenges. These directions include (i) using more hardware-related information from the mobile device's camera and other available sensors, (ii) capturing high-quality data to reflect real-world AR scenarios, and (iii) designing a model architecture to utilize the new information.

Tian Guo, Hui Yuan, Qi Liu et al.

Point cloud compression significantly reduces data volume but sacrifices reconstruction quality, highlighting the need for advanced quality enhancement techniques. Most existing approaches focus primarily on point-to-point fidelity, often neglecting the importance of perceptual quality as interpreted by the human visual system. To address this issue, we propose a generative adversarial network for point cloud quality enhancement (PCE-GAN), grounded in optimal transport theory, with the goal of simultaneously optimizing both data fidelity and perceptual quality. The generator consists of a local feature extraction (LFE) unit, a global spatial correlation (GSC) unit and a feature squeeze unit. The LFE unit uses dynamic graph construction and a graph attention mechanism to efficiently extract local features, placing greater emphasis on points with severe distortion. The GSC unit uses the geometry information of neighboring patches to construct an extended local neighborhood and introduces a transformer-style structure to capture long-range global correlations. The discriminator computes the deviation between the probability distributions of the enhanced point cloud and the original point cloud, guiding the generator to achieve high quality reconstruction. Experimental results show that the proposed method achieves state-of-the-art performance. Specifically, when applying PCE-GAN to the latest geometry-based point cloud compression (G-PCC) test model, it achieves an average BD-rate of -19.2% compared with the PredLift coding configuration and -18.3% compared with the RAHT coding configuration. Subjective comparisons show a significant improvement in texture clarity and color transitions, revealing finer details and more natural color gradients.

Tian Guo, Chen Chen, Hui Yuan et al.

Point cloud sampling plays a crucial role in reducing computation costs and storage requirements for various vision tasks. Traditional sampling methods, such as farthest point sampling, lack task-specific information and, as a result, cannot guarantee optimal performance in specific applications. Learning-based methods train a network to sample the point cloud for the targeted downstream task. However, they do not guarantee that the sampled points are the most relevant ones. Moreover, they may result in duplicate sampled points, which requires completion of the sampled point cloud through post-processing techniques. To address these limitations, we propose a contribution-based sampling network (CS-Net), where the sampling operation is formulated as a Top-k operation. To ensure that the network can be trained in an end-to-end way using gradient descent algorithms, we use a differentiable approximation to the Top-k operation via entropy regularization of an optimal transport problem. Our network consists of a feature embedding module, a cascade attention module, and a contribution scoring module. The feature embedding module includes a specifically designed spatial pooling layer to reduce parameters while preserving important features. The cascade attention module combines the outputs of three skip connected offset attention layers to emphasize the attractive features and suppress less important ones. The contribution scoring module generates a contribution score for each point and guides the sampling process to prioritize the most important ones. Experiments on the ModelNet40 and PU147 showed that CS-Net achieved state-of-the-art performance in two semantic-based downstream tasks (classification and registration) and two reconstruction-based tasks (compression and surface reconstruction).

Yiqin Zhao, Mallesham Dasari, Tian Guo

High-quality environment lighting is essential for creating immersive mobile augmented reality (AR) experiences. However, achieving visually coherent estimation for mobile AR is challenging due to several key limitations in AR device sensing capabilities, including low camera FoV and limited pixel dynamic ranges. Recent advancements in generative AI, which can generate high-quality images from different types of prompts, including texts and images, present a potential solution for high-quality lighting estimation. Still, to effectively use generative image diffusion models, we must address two key limitations of content quality and slow inference. In this work, we design and implement a generative lighting estimation system called CleAR that can produce high-quality, diverse environment maps in the format of 360° HDR images. Specifically, we design a two-step generation pipeline guided by AR environment context data to ensure the output aligns with the physical environment's visual context and color appearance. To improve the estimation robustness under different lighting conditions, we design a real-time refinement component to adjust lighting estimation results on AR devices. Through a combination of quantitative and qualitative evaluations, we show that CleAR outperforms state-of-the-art lighting estimation methods on both estimation accuracy, latency, and robustness, and is rated by 31 participants as producing better renderings for most virtual objects. For example, CleAR achieves 51% to 56% accuracy improvement on virtual object renderings across objects of three distinctive types of materials and reflective properties. CleAR produces lighting estimates of comparable or better quality in just 3.2 seconds -- over 110X faster than state-of-the-art methods.

Yiyang Zhao, Yunzhuo Liu, Bo Jiang et al.

This work presents a novel approach to neural architecture search (NAS) that aims to increase carbon efficiency for the model design process. The proposed framework CE-NAS addresses the key challenge of high carbon cost associated with NAS by exploring the carbon emission variations of energy and energy differences of different NAS algorithms. At the high level, CE-NAS leverages a reinforcement-learning agent to dynamically adjust GPU resources based on carbon intensity, predicted by a time-series transformer, to balance energy-efficient sampling and energy-intensive evaluation tasks. Furthermore, CE-NAS leverages a recently proposed multi-objective optimizer to effectively reduce the NAS search space. We demonstrate the efficacy of CE-NAS in lowering carbon emissions while achieving SOTA results for both NAS datasets and open-domain NAS tasks. For example, on the HW-NasBench dataset, CE-NAS reduces carbon emissions by up to 7.22X while maintaining a search efficiency comparable to vanilla NAS. For open-domain NAS tasks, CE-NAS achieves SOTA results with 97.35% top-1 accuracy on CIFAR-10 with only 1.68M parameters and a carbon consumption of 38.53 lbs of CO2. On ImageNet, our searched model achieves 80.6% top-1 accuracy with a 0.78 ms TensorRT latency using FP16 on NVIDIA V100, consuming only 909.86 lbs of CO2, making it comparable to other one-shot-based NAS baselines.

Yiyang Zhao, Linnan Wang, Kevin Yang et al.

In contrast to single-objective optimization (SOO), multi-objective optimization (MOO) requires an optimizer to find the Pareto frontier, a subset of feasible solutions that are not dominated by other feasible solutions. In this paper, we propose LaMOO, a novel multi-objective optimizer that learns a model from observed samples to partition the search space and then focus on promising regions that are likely to contain a subset of the Pareto frontier. The partitioning is based on the dominance number, which measures "how close" a data point is to the Pareto frontier among existing samples. To account for possible partition errors due to limited samples and model mismatch, we leverage Monte Carlo Tree Search (MCTS) to exploit promising regions while exploring suboptimal regions that may turn out to contain good solutions later. Theoretically, we prove the efficacy of learning space partitioning via LaMOO under certain assumptions. Empirically, on the HyperVolume (HV) benchmark, a popular MOO metric, LaMOO substantially outperforms strong baselines on multiple real-world MOO tasks, by up to 225% in sample efficiency for neural architecture search on Nasbench201, and up to 10% for molecular design.

Yiqin Zhao, Tian Guo

Omnidirectional lighting provides the foundation for achieving spatially-variant photorealistic 3D rendering, a desirable property for mobile augmented reality applications. However, in practice, estimating omnidirectional lighting can be challenging due to limitations such as partial panoramas of the rendering positions, and the inherent environment lighting and mobile user dynamics. A new opportunity arises recently with the advancements in mobile 3D vision, including built-in high-accuracy depth sensors and deep learning-powered algorithms, which provide the means to better sense and understand the physical surroundings. Centering the key idea of 3D vision, in this work, we design an edge-assisted framework called Xihe to provide mobile AR applications the ability to obtain accurate omnidirectional lighting estimation in real time. Specifically, we develop a novel sampling technique that efficiently compresses the raw point cloud input generated at the mobile device. This technique is derived based on our empirical analysis of a recent 3D indoor dataset and plays a key role in our 3D vision-based lighting estimator pipeline design. To achieve the real-time goal, we develop a tailored GPU pipeline for on-device point cloud processing and use an encoding technique that reduces network transmitted bytes. Finally, we present an adaptive triggering strategy that allows Xihe to skip unnecessary lighting estimations and a practical way to provide temporal coherent rendering integration with the mobile AR ecosystem. We evaluate both the lighting estimation accuracy and time of Xihe using a reference mobile application developed with Xihe's APIs. Our results show that Xihe takes as fast as 20.67ms per lighting estimation and achieves 9.4% better estimation accuracy than a state-of-the-art neural network.

Shijian Li, Oren Mangoubi, Lijie Xu et al.

Stochastic Gradient Descent (SGD) has become the de facto way to train deep neural networks in distributed clusters. A critical factor in determining the training throughput and model accuracy is the choice of the parameter synchronization protocol. For example, while Bulk Synchronous Parallel (BSP) often achieves better converged accuracy, the corresponding training throughput can be negatively impacted by stragglers. In contrast, Asynchronous Parallel (ASP) can have higher throughput, but its convergence and accuracy can be impacted by stale gradients. To improve the performance of synchronization protocol, recent work often focuses on designing new protocols with a heavy reliance on hard-to-tune hyper-parameters. In this paper, we design a hybrid synchronization approach that exploits the benefits of both BSP and ASP, i.e., reducing training time while simultaneously maintaining the converged accuracy. Based on extensive empirical profiling, we devise a collection of adaptive policies that determine how and when to switch between synchronization protocols. Our policies include both offline ones that target recurring jobs and online ones for handling transient stragglers. We implement the proposed policies in a prototype system, called Sync-Switch, on top of TensorFlow, and evaluate the training performance with popular deep learning models and datasets. Our experiments show that Sync-Switch achieves up to 5.13X throughput speedup and similar converged accuracy when comparing to BSP. Further, we observe that Sync-Switch achieves 3.8% higher converged accuracy with just 1.23X the training time compared to training with ASP. Moreover, Sync-Switch can be used in settings when training with ASP leads to divergence errors. Sync-Switch achieves all of these benefits with very low overhead, e.g., the framework overhead can be as low as 1.7% of the total training time.

Yiyang Zhao, Linnan Wang, Yuandong Tian et al.

Efficient evaluation of a network architecture drawn from a large search space remains a key challenge in Neural Architecture Search (NAS). Vanilla NAS evaluates each architecture by training from scratch, which gives the true performance but is extremely time-consuming. Recently, one-shot NAS substantially reduces the computation cost by training only one supernetwork, a.k.a. supernet, to approximate the performance of every architecture in the search space via weight-sharing. However, the performance estimation can be very inaccurate due to the co-adaption among operations. In this paper, we propose few-shot NAS that uses multiple supernetworks, called sub-supernet, each covering different regions of the search space to alleviate the undesired co-adaption. Compared to one-shot NAS, few-shot NAS improves the accuracy of architecture evaluation with a small increase of evaluation cost. With only up to 7 sub-supernets, few-shot NAS establishes new SoTAs: on ImageNet, it finds models that reach 80.5% top-1 accuracy at 600 MB FLOPS and 77.5% top-1 accuracy at 238 MFLOPS; on CIFAR10, it reaches 98.72% top-1 accuracy without using extra data or transfer learning. In Auto-GAN, few-shot NAS outperforms the previously published results by up to 20%. Extensive experiments show that few-shot NAS significantly improves various one-shot methods, including 4 gradient-based and 6 search-based methods on 3 different tasks in NasBench-201 and NasBench1-shot-1.

Shijian Li, Robert J. Walls, Tian Guo

Cloud GPU servers have become the de facto way for deep learning practitioners to train complex models on large-scale datasets. However, it is challenging to determine the appropriate cluster configuration---e.g., server type and number---for different training workloads while balancing the trade-offs in training time, cost, and model accuracy. Adding to the complexity is the potential to reduce the monetary cost by using cheaper, but revocable, transient GPU servers. In this work, we analyze distributed training performance under diverse cluster configurations using CM-DARE, a cloud-based measurement and training framework. Our empirical datasets include measurements from three GPU types, six geographic regions, twenty convolutional neural networks, and thousands of Google Cloud servers. We also demonstrate the feasibility of predicting training speed and overhead using regression-based models. Finally, we discuss potential use cases of our performance modeling such as detecting and mitigating performance bottlenecks.

Yiqin Zhao, Tian Guo

We propose an efficient lighting estimation pipeline that is suitable to run on modern mobile devices, with comparable resource complexities to state-of-the-art mobile deep learning models. Our pipeline, PointAR, takes a single RGB-D image captured from the mobile camera and a 2D location in that image, and estimates 2nd order spherical harmonics coefficients. This estimated spherical harmonics coefficients can be directly utilized by rendering engines for supporting spatially variant indoor lighting, in the context of augmented reality. Our key insight is to formulate the lighting estimation as a point cloud-based learning problem directly from point clouds, which is in part inspired by the Monte Carlo integration leveraged by real-time spherical harmonics lighting. While existing approaches estimate lighting information with complex deep learning pipelines, our method focuses on reducing the computational complexity. Through both quantitative and qualitative experiments, we demonstrate that PointAR achieves lower lighting estimation errors compared to state-of-the-art methods. Further, our method requires an order of magnitude lower resource, comparable to that of mobile-specific DNNs.

Matthew LeMay, Shijian Li, Tian Guo

Deep learning models are increasingly used for end-user applications, supporting both novel features such as facial recognition, and traditional features, e.g. web search. To accommodate high inference throughput, it is common to host a single pre-trained Convolutional Neural Network (CNN) in dedicated cloud-based servers with hardware accelerators such as Graphics Processing Units (GPUs). However, GPUs can be orders of magnitude more expensive than traditional Central Processing Unit (CPU) servers. These resources could also be under-utilized facing dynamic workloads, which may result in inflated serving costs. One potential way to alleviate this problem is by allowing hosted models to share the underlying resources, which we refer to as multi-tenant inference serving. One of the key challenges is maximizing the resource efficiency for multi-tenant serving given hardware with diverse characteristics, models with unique response time Service Level Agreement (SLA), and dynamic inference workloads. In this paper, we present Perseus, a measurement framework that provides the basis for understanding the performance and cost trade-offs of multi-tenant model serving. We implemented Perseus in Python atop a popular cloud inference server called Nvidia TensorRT Inference Server. Leveraging Perseus, we evaluated the inference throughput and cost for serving various models and demonstrated that multi-tenant model serving led to up to 12% cost reduction.

Peter M. VanNostrand, Ioannis Kyriazis, Michelle Cheng et al.

Performing deep learning on end-user devices provides fast offline inference results and can help protect the user's privacy. However, running models on untrusted client devices reveals model information which may be proprietary, i.e., the operating system or other applications on end-user devices may be manipulated to copy and redistribute this information, infringing on the model provider's intellectual property. We propose the use of ARM TrustZone, a hardware-based security feature present in most phones, to confidentially run a proprietary model on an untrusted end-user device. We explore the limitations and design challenges of using TrustZone and examine potential approaches for confidential deep learning within this environment. Of particular interest is providing robust protection of proprietary model information while minimizing total performance overhead.

Shijian Li, Robert J. Walls, Lijie Xu et al.

Distributed training frameworks, like TensorFlow, have been proposed as a means to reduce the training time of deep learning models by using a cluster of GPU servers. While such speedups are often desirable---e.g., for rapidly evaluating new model designs---they often come with significantly higher monetary costs due to sublinear scalability. In this paper, we investigate the feasibility of using training clusters composed of cheaper transient GPU servers to get the benefits of distributed training without the high costs. We conduct the first large-scale empirical analysis, launching more than a thousand GPU servers of various capacities, aimed at understanding the characteristics of transient GPU servers and their impact on distributed training performance. Our study demonstrates the potential of transient servers with a speedup of 7.7X with more than 62.9% monetary savings for some cluster configurations. We also identify a number of important challenges and opportunities for redesigning distributed training frameworks to be transient-aware. For example, the dynamic cost and availability characteristics of transient servers suggest the need for frameworks to dynamically change cluster configurations to best take advantage of current conditions.

Tian Guo

Modern mobile applications are benefiting significantly from the advancement in deep learning, e.g., implementing real-time image recognition and conversational system. Given a trained deep learning model, applications usually need to perform a series of matrix operations based on the input data, in order to infer possible output values. Because of computational complexity and size constraints, these trained models are often hosted in the cloud. To utilize these cloud-based models, mobile apps will have to send input data over the network. While cloud-based deep learning can provide reasonable response time for mobile apps, it restricts the use case scenarios, e.g. mobile apps need to have network access. With mobile specific deep learning optimizations, it is now possible to employ on-device inference. However, because mobile hardware, such as GPU and memory size, can be very limited when compared to its desktop counterpart, it is important to understand the feasibility of this new on-device deep learning inference architecture. In this paper, we empirically evaluate the inference performance of three Convolutional Neural Networks (CNNs) using a benchmark Android application we developed. Our measurement and analysis suggest that on-device inference can cost up to two orders of magnitude greater response time and energy when compared to cloud-based inference, and that loading model and computing probability are two performance bottlenecks for on-device deep inferences.